Study the val_loss of DCAE depending on its window and strize parameters

[vrodriguezf]

In GitLab by @lgs on Oct 29, 2020, 10:49

null

[vrodriguezf]

In GitLab by @lgs on Oct 29, 2020, 10:49

assigned to @lgs

[vrodriguezf]

In GitLab by @lgs on Oct 29, 2020, 11:09

I have updated the 02_DCAE to train it with a papermill experiment and a wandb sweep. I have included the following code to read the parameters from _experiments_papermill_caler.ipynb

ifnone(config.get('variable'), default_value)

I have launched a training sweep to study the val_loss of autoencoders with all possible combinations of the windows_size (range(24,144,12)) and stride (range(1,12,2)) parameters.

[vrodriguezf]

In GitLab by @lgs on Oct 29, 2020, 12:35

After the first sweep, we got 60 DCAE models available here.

When we take the results with a 'val_loss' lower than 0.3, we get a set of 14 models. In these models, it seems that the size of the window is more important to obtain a good-quality autoencoder than the stride. In the graph, we can see that for the models with a val_loss<=0.30, most of the models used windows below 60. In the case of the stride there is not such a clear pattern, but it seems that low strides (high redundancy in the windowing) improves the quality of the models.

I will re-launch a sweep in which the models are calculated with a windows size range(24,72,12) and a stride in individual steps between 1 and 9. To check in more detail what is happening.

[vrodriguezf]

Sweeps should be redone, closing